A briefing at every start
Facts, decisions, failed attempts, preferences and the last session’s open steps, ranked by what has proven useful, within a token budget.
agent-mem records what your coding agents do, writes short facts from it in the background, and
reminds every agent of what matters at the right moment. Nothing leaves your computer. Everything sits in one
SQLite file, and you can browse it at http://localhost:37778.
Install the command (macOS, Linux, Windows):
uv tool install "agent-mem[semantic,observer] @ git+https://github.com/Abuarchiv/agent-mem"Connect every agent installed on this machine:
agent-mem setup all --writeCheck that it works:
agent-mem doctorStart a session in any repository and send one request. The next session begins with what the first one learned.
A briefing at every start
Facts, decisions, failed attempts, preferences and the last session’s open steps, ranked by what has proven useful, within a token budget.
The right memory at the right moment
Matching memories for a prompt, only when the match is strong. The history of a file before it is read or edited. How an error was fixed last time.
Corrections that stick
Say “use pnpm instead of npm” twice and every agent is told before it runs npm. Switch the rule on and the
command is blocked.
One memory for every agent
What Claude Code learned is there for Codex, Copilot and OpenCode, and the other way round.
Facts without an API key
The CLI you already pay for writes the facts in the background: Claude Haiku for Claude Code, GPT-6-Luna for Codex and Copilot, the free Big Pickle for OpenCode. A request never leaves its vendor.
Local and private
No telemetry, no cloud, no server. Secrets are removed before storage; <private> text is never stored.
| Moment | What happens |
|---|---|
| Session start | A briefing for the agent, a panel for you: project memory, which model writes facts, recent work, what needs your attention |
| You send a request | Matching memories, only when the match is strong; the terminal stays quiet |
| Before a command or edit | A note if it matches something you corrected, the file’s history, or a block for a rule you switched on |
| A command fails | If this error was fixed before, the agent sees how |
| The agent finishes | The request is saved; a background process writes facts, a title, the session summary and lessons |
| Once a day | Duplicates and replaced facts are retired; your own notes are never touched |
| Compaction | Session goal, decisions and open problems are given back to the agent |
| Session end | What was saved: requests, files, errors |
Read Architecture Overview for the full picture.
| Agents | Claude Code 2.x, Codex CLI 0.x, GitHub Copilot CLI 1.0.103, the Copilot app 1.1.29 (macOS), Copilot Chat in VS Code 1.142, OpenCode 1.19 and 2.2 |
| Python | 3.11 to 3.14 |
| OS | macOS (arm64; Intel without search by meaning), Linux, Windows |
| Install size | about 10 MB, plus 220 MB for search by meaning, plus about 400 MB for background facts |